Semantic Segmentation Method Based on Multiscale Feature Alignment and Aggregation

During semantic segmentation of images, a convolutional neural network easily misplaces the high-level features with low-level features after down-sampling and padding operations. To solve the mismatch problem between high- and low-level features and better aggregate the multiscale feature information, this paper proposes a semantic segmentation method with a multiscale feature alignment aggregation (MFAA) module. The MFAA module adopts a learnable interpolation strategy to learn pixel transform migration, thereby alleviating the feature-misalignment problem of feature aggregation at different scales. The module includes an attention mechanism that improves the decoder's ability to recover the important details. Using multiple MFAA modules, the semantic information of high-level features, and the spatial information of low-level features, this method aligns and aggregates the high- and low-level features to refine the semantic segmentation effect. The proposed network structure was validated on PASCAL VOC 2012. Using a ResNet- 50 backbone network, the mean intersection-over-union reached 78. 4% on the validation set. Experimentally, the proposed method achieved better evaluation indices than several mainstream segmentation methods and effectively improved the image segmentation effect.